diff --git a/optimum/exporters/executorch/recipes/metal.py b/optimum/exporters/executorch/recipes/metal.py
index 0ca1c56..cf7fe8b 100644
--- a/optimum/exporters/executorch/recipes/metal.py
+++ b/optimum/exporters/executorch/recipes/metal.py
@@ -31,6 +31,8 @@
         METAL_BACKEND_AVAILABLE = False
 
 if METAL_BACKEND_AVAILABLE:
+    import torch
+
     from tabulate import tabulate
     from torch.export import ExportedProgram
 
@@ -52,6 +54,13 @@
     )
     from ..recipe_registry import register_recipe
 
+    def _linear_bias_decomposition(input, weight, bias=None):
+        weight_t = torch.ops.aten.t.default(weight)
+        out = torch.ops.aten.matmul.default(input, weight_t)
+        if bias is not None:
+            return torch.ops.aten.add.Tensor(out, bias)
+        return out
+
     @register_recipe("metal")
     def export_to_executorch_with_metal(
         model: Union[
@@ -89,6 +98,13 @@ def _lower_to_executorch(
             if len(exported_programs) == 1:
                 exported_programs = {"forward": next(iter(exported_programs.values()))}
 
+            # Decompose linear+bias into matmul+add to avoid addmm,
+            # which the Metal backend doesn't support.
+            for key in exported_programs:
+                decomp_table = torch.export.default_decompositions()
+                decomp_table[torch.ops.aten.linear.default] = _linear_bias_decomposition
+                exported_programs[key] = exported_programs[key].run_decompositions(decomp_table)
+
             partitioners = {
                 key: [MetalPartitioner([MetalBackend.generate_method_name_compile_spec(key)])]
                 for key in exported_programs.keys()